Okay, looks like we've tracked down the problem -- it seems to be a 
misconfiguration of the Cisco box. (Moral: Never trust a Cisco intern to 
configure your Cisco box ;-) )

Thanks for your help tracking this down,
Glen



Glen Gibb wrote:
> Ah, I'd completely missed that the ARPs were wrapped in ISL :-( I'd seen 
> that for some of the others packet types...
>
> I'll post an update later about progress at this end.
>
> Glen
>
>
> Natasha Gude wrote:
>   
>> Hey Glen,
>>
>> So the reason I didn't see the ARPs was because I was using tcpdump 
>> and they came up as SNAP packets.  Opening it up in wireshark, it 
>> turns out all of the ARPs are being encapsulated by Cisco's ISL 
>> protocol.  So I think what's happening is that mvm-17 and mvm-33 are 
>> sending out ARPs, which the 6k is encapsulating in ISL, which is then 
>> getting sent up the controller.  These packets are getting flooded 
>> out, to all of the Openflow ports, including ones not connected to the 
>> 6k, which I'm guessing would have stripped that outer header.  Are 
>> mvm-44 and 24 on eth2 and eth3?  Are they perhaps receiving these 
>> packets and not knowing how to respond?
>>
>> So it seems like this might be a 6k configuration problem rather than 
>> an Openflow/NOX one, but let me know if you have more information.
>>
>> Also, it's probably worth noting that mvm-17 and mvm-33 seem to have 
>> the same mac address (probably because they're vms?).  When the ISL 
>> stuff gets figured out, I'm guessing this will cause some issues.
>>
>> Natasha
>>
>> Glen Gibb wrote:
>>     
>>> Thanks for getting back to me Natasha,
>>>
>>>
>>> Natasha Gude wrote:
>>>       
>>>> Hey Glen,
>>>>
>>>> First off, thanks for all of the debugging output - it really helped 
>>>> in figuring out your setup.
>>>>
>>>>         
>>> No probs. I figure the more useful information I give you the more 
>>> likely you are to be able to help with the problem. Let me correct 
>>> one minor mistake in the previous e-mail. It appears that mvm-17 is 
>>> on eth1 and mvm-33 is on eth4 (both via Cat6k) -- I had said they 
>>> were both on eth1.
>>>
>>>
>>>       
>>>> I took a look at the dump files and nox log, and I didn't see more 
>>>> than a few ARP packets, but I did see a lot of cisco traffic, 
>>>> presumably from the cat6k.  I'm not sure if there's necessarily a 
>>>> problem here, but I can tell you what it seems like the situation is 
>>>> and you can let me know if NOX should be behaving differently.
>>>>         
>>> The ARP traffic I was referring to can be seen from packet 212 on in 
>>> b.eth1.dump and packet 199 in b.eth2.dump (and similar locations in 
>>> the other two). Once that first ARP request comes in they appear on 
>>> the order of < 100ms :-(
>>>
>>>
>>>       
>>>> Almost all of the flows seen by NOX are from the mac 
>>>> 00:01:63:d4:67:ca (which I'm assuming is the 6k).  Because the 6k is 
>>>> connected to two of the Openflow switch's ports, NOX has record of 
>>>> the mac at these two locations (eth1 and eth4, which are port 
>>>> numbers 0 and 3 respectively in the NOX log file).  Right now, we 
>>>> have a notion of a "primary" location when a sender is connected at 
>>>> two different points in the network.  At any given time, a sender's 
>>>> primary location is the location it most recently sent a packet 
>>>> from.  When that location switches, the old one is "poisoned" to 
>>>> force ongoing traffic to be routed to the new location.  Thus the 
>>>> poison dbg message you were seeing results from the 6k sending a 
>>>> packet with that source mac address through a different port on the 
>>>> openflow switch than which it last sent through.  What's interesting 
>>>> is that to begin with, a different mac address is used by the 6k 
>>>> when sending traffic to openflow's eth4 interface 
>>>> (00:01:63:d4:67:cb), but then when the destination address changes 
>>>> to 01:00:0c:00:00:07, the source address is always 00:01:63:d4:67:ca 
>>>> regardless of the openflow port it is received on.
>>>>         
>>> Actually all traffic from the Cisco on eth4 is from 
>>> 00:01:63:d4:67:cb. We only start seeing things listed as being from 
>>> the :ca address on eth4 when NOX is running.
>>>
>>>
>>>       
>>>> The second point worth noting is that all of the 6k's traffic is 
>>>> sent to multicast addresses, and NOX currently treats multicast and 
>>>> broadcast traffic the same, flooding a packet out every port except 
>>>> for the one it came in on.  If there's a more appropriate way of 
>>>> dealing with this traffic, please let me know!
>>>>         
>>> As far as I know it should be okay to simply flood this traffic.
>>>
>>>       
>>>> So that's what seems to be the situation.  Again, let me know if you 
>>>> think any of the above described behavior is incorrect, or if 
>>>> there's still a problem that I just couldn't deduce from the 
>>>> log/dump files I looked at.
>>>>         
>>> Glen
>>>       
>>     
>
>
> _______________________________________________
> nox-dev mailing list
> [email protected]
> http://noxrepo.org/mailman/listinfo/nox-dev_noxrepo.org
>
>   


_______________________________________________
nox-dev mailing list
[email protected]
http://noxrepo.org/mailman/listinfo/nox-dev_noxrepo.org

Reply via email to